The readiness assessment has become a standard opening move, and a largely ceremonial one. It produces a maturity score against someone’s five-level model, a roadmap with three horizons, and a list of use cases ranked by an impact-versus-effort matrix that nobody revisits. It rarely answers the question the sponsor actually has, which is whether the first thing we build will work.
A useful assessment answers three concrete questions instead. Is the data an agent would need available at the moment the decision is made, or only in a warehouse the following morning? Does a human currently make this decision with information the system holds, or with information they carry in their head? And when the agent is wrong — which it will be — what does the organisation lose, and who notices?
Those questions sort candidate use cases far more usefully than impact and effort, because they identify the ones that fail for reasons no amount of model quality can fix. A purchase-approval agent in a business where the real approval logic lives in a WhatsApp group is not an AI problem. Finding that out in an assessment costs a fortnight; finding it out in a pilot costs a quarter and some credibility.
This is what the Crawl, Walk, Run sequence is for, and why the order matters. Crawl establishes that the data supports the decision and that someone owns the outcome. Walk puts the agent in the path with a human confirming its output, which produces the labelled evidence the business case needs. Run removes the confirmation step for the cases where the evidence supports it — and only those cases.
The deliverable to insist on is not a maturity score. It is a shortlist of decisions where the data is present, the owner is named, and the cost of being wrong is understood — plus an honest list of the ones that are not ready and what would make them so.
